[Measurement] Add barrel + factory fixtures to extension-benchmark corpus

[AndresL230]

Summary

PR #110 (A3 barrel re-exports + A5 factory/DI/aliased clients) added detection coverage for several cross-file patterns, but the D1 benchmark corpus in recost-dev/extension-benchmark does not exercise any of them. The recall gain (+2.94pp on the global metric) is real but not measurable per-pattern, and B2 (#81) dual-location work needs these fixtures to be testable.

Fixtures to add (in recost-dev/extension-benchmark)

1. barrel-direct-reexport — import { X } from "./barrel" where barrel.ts does export { X } from "./real".
2. barrel-aliased-reexport — export { x as y } from "./real".
3. barrel-wildcard — export * from "./real".
4. barrel-default — export { default } from "./real".
5. factory-in-file — export function makeClient() { return new OpenAI(); } consumed in same file.
6. factory-cross-file — same as 5 but consumer is a sibling file.
7. di-typed-constructor — class C { constructor(private ai: OpenAI) {} useIt() { this.ai.method(); } }.

Each fixture: src/, expected.json listing the resolved provider, file:line of each call site.

What to do here (in recost-dev/extension)

• Once fixtures land in extension-benchmark, refresh benchmark/baseline.json to capture the new per-fixture metrics.
• No code changes expected — A3/A5 already handles these patterns.

Acceptance criteria

• 7 new fixtures in extension-benchmark covering each barrel/factory/DI shape.
• Per-fixture metrics in benchmark output for each.
• baseline.json refreshed to include them.

Depends on

• PR fix(detection): A3 + A5 — barrel re-exports + factory/DI/aliased clients (closes #75, #77) #110 (already merged) — defines what should resolve.

Blocks

• [Traceability] Dual locations for cross-file resolved calls (B2) #81 (B2) — dual-location accuracy can only be measured on fixtures that exercise cross-file resolution.

Reference

Memory: a3_a5_pr_status.md ("Corpus gap").